Skip to content

Add Gemma4 EAGLE-3 hidden-state taps#20150

Draft
digantdesai wants to merge 1 commit into
gh/digantdesai/54/headfrom
gh/digantdesai/55/head
Draft

Add Gemma4 EAGLE-3 hidden-state taps#20150
digantdesai wants to merge 1 commit into
gh/digantdesai/54/headfrom
gh/digantdesai/55/head

Conversation

@digantdesai

Copy link
Copy Markdown
Contributor

Adds validated tap-layer configuration, a shared decode path that can
return concatenated tap features, and forward_logits_taps for
soft-capped logits with EAGLE-3 features. Includes tests for tap
ordering, validation, logits parity, and last-logits behavior.

Authored with assistance from Claude Code.

[ghstack-poisoned]
@pytorch-bot

pytorch-bot Bot commented Jun 9, 2026

Copy link
Copy Markdown

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20150

Note: Links to docs will display an error until the docs builds have been completed.

⏳ No Failures, 1 Pending

As of commit 57ef8a5 with merge base dc55469 (image):
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

self,
tokens: torch.LongTensor,
input_pos: torch.LongTensor,
last_logits_only: bool = True,

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

For tapped model, do we need all logits, not just final so sampling can be done outside model?

If so, why is last_logits_only default to True on forward_logits_taps?

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants